Events

Leif E. Peterson, Ph.D.: MPH-Multiscale Machine Learning for Diagnostic Class Discovery and Prediction

This presentation will journey through classical statistical techniques and more contemporary works involving computational intelligence via neural adaptive learning and metaheuristics for the purpose of solving complex problems based on low-to-high dimensional data. Multiple scales will be discussed in the context of drawing information at high levels of granularity up to the psychosocial scale, which influences health behavior and risk taking. Class discovery will be introduced as a way to first understand the cluster structure of datasets based on newly spawned novel data independent from known diagnostic class.

Concept clustering hinged to text mining will also be described. Once cluster structures and concepts are identified, cross validation can be pursued for feature selection, followed by identification of linear separability. Prediction of diagnostic class of future patients can finally be addressed with the use of cross-validation during crisp and fuzzy classification analysis. The pendulum paradox of new technology will be briefly discussed in order to support data integration using non-linear manifold learning for dimension reduction and decorrelation with fast wavelet transformations for removing noninformative data.

Additionally, multivariate component subtraction based on the Marčenko-Pastur law will be discussed for exploiting the signal-noise cutoff in a dataset for information retrieval.

Back to top